Nov 2–8, 2026

Claude Code 2.1.319 changes how MCP tasks run and how auto mode handles actions it cannot check server-side.

Jaquatech released AgentDesk 1.0, a local development environment that assigns coding work to planner, developer and reviewer agents, with a human pressing Merge.

Artificial Analysis has launched VIALS-AA, an independent evaluation of language models on 161 scientific-image interpretation tasks.

Anthropic is gradually merging Claude Cowork and chat into one conversation experience across web, desktop, and mobile.

AWS describes a sample that runs Claude Code tasks in isolated Amazon Bedrock AgentCore sessions and returns GitHub pull requests for human review.

Facade Code presents a VS Code extension for running native coding-agent CLIs alongside project files and coordinating agents through messaging and terminal control.

Gauge has open-sourced an MIT-licensed CLI that runs coding agents against a product in local Docker containers.

Greenlight Financial says its Luna system resolves more than 90% of up to 4,000 monthly security and risk cases without human intervention, using a tiered multi-agent architecture on Amazon Bedrock AgentCore.

A research team introduced InteractiveWorld, a simulated environment for testing multimodal assistants that guide people through physical tasks, and InteractiveBench, a 120-task benchmark built on it.

Joël Niklaus’s Harness Leaderboard compares nine coding-agent harnesses while keeping model, provider, and reasoning effort fixed.

llamawatch is a self-hosted dashboard for monitoring local language-model servers, hardware, jobs, spending and agent runs from one interface.

Mozilla AI has announced Otari Subscriptions, a monthly plan that bundles access to open-weight models, provider routing and coding-agent tools behind one key.

Rabbit v0.0.1 is a Claude plugin that turns unfinished Jira sprint tasks into separate, staged coding sessions, with human approvals, an independent read-only reviewer and explicit integration-test results.

BenBish’s open-source resume-plugin-cc moves unfinished Codex or Grok coding work into a fresh Claude Code or Devin session.

Artificial Analysis has published Stirrup, an MIT-licensed Python framework and starting template for building agents with tools, sessions, and configurable model clients.

ThruLine’s 0.3.0 Mac app checks coding-agent claims against locally recorded session logs, including what ran and on which code version.

AI Perceivable’s apexe 0.10.0 adds a shell launcher and setup commands for Claude Code, Codex, OpenCode and pi.

Anthropic is rolling out Personal Finance in Claude to Pro, Max and Ultra subscribers in the US.

Anthropic merged a pull request publishing the source of Claude Code’s built-in privacy mod under mods/privacy.

Anthropic’s Claude Code v2.1.319 release adds terminal support for MCP tasks that continue in the background, plus a permission setting that denies calls that would otherwise wait for a prompt.

Claude Desktop’s October 20, 2026 changelog added a Storage panel and a /storage command for checking the app’s local disk use and clearing some of it.

Claude Code Game Studios is an open-source template that organizes a Claude Code session into a game-development workflow with specialized agents, skills, hooks and configurable quality checks.

A team of researchers introduces VISTA, a visual harness that lets a multimodal model retain and retrieve original observations while working through interactive environments.

Benchling researchers introduce BenchBench-Protocol, a benchmark of 149 wet-lab protocol modification tasks drawn from changes scientists made during real experimental work.

EnvOpt is a research method for automatically optimizing the information environment in which an agent operates, alongside its model and harness.

A study of 36,710 software repositories found autonomous coding-agent loops in 217, while finding little repository evidence of the persistent state files practitioners recommend.

Dwarkesh Patel built a Blood on the Clocktower environment to compare how AI models coordinate, deceive and reason under changing, incomplete information.

Ai2 introduced ToyBlocks, an open simulation framework for studying robot learning with a tabletop robot arm and blocks.

Surge AI compared four Claude Code harness configurations on 100 software-engineering tasks and found that raising Opus 5.5's thinking effort produced a larger score gain than any tested harness addition.

Anthropic is gradually rolling out Personal Finance in Claude to US subscribers on Pro, Max, and Ultra plans.

Claude Code 2.1.317 changes how its game-cloning rule handles requests that mention an existing title.

A developer used Claude Code Mods to change the existing /copy command so copied assistant text includes the model name and an attribution wrapper.

A Washington Post investigation, summarized by Ben Werdmuller, found that an AI-generated election-information site called Ballot Brief appeared frequently in chatbot answers about the US election.

Conductor says it is joining Meta and has set a timetable for winding down its cloud service.

Anthropic’s Claude Cookbook shows how to build a background workflow that turns completed Claude Managed Agents sessions into a linked, searchable memory store for future runs.

GradientDeck offers 1,286 DESIGN.md files based on real websites, which users can browse and copy into a repository for Claude Code, Codex or Cursor.

Firecrawl is presenting its Alexandria Research Index as a free source of life-science papers for agents in ChatGPT, Claude and Grok.

Fireworks says Moonshot AI’s Kimi K3.1 Pro is now available through its serverless inference service, with token pricing and an option for dedicated deployment.

Language models can spend too much effort on easy questions and too little on steps that need checking.

Conductor says it is joining Meta and will wind down its cloud service: Conductor Cloud becomes read-only on November 13 and shuts down on November 20.

Anthropic is rolling out Personal Finance in Claude, a U.S.

Apple has reportedly lowered daily AI spending caps for some employees, with further cuts under consideration across engineering and other teams.

Zed Industries’ October update to Delta adds a Claude Code plugin for sharing agent sessions, remote MCP server connections, and deeper subthreads.

Agno’s October 2026 roundup describes new ways to run Claude Code and Codex as managed agents, including resumable sandboxes, background runs, permission approvals, and persistent files.

BioResearchBench tests frontier models on 50 expert-developed biological research questions spanning 10 subdomains and multiple data types.

Anthropic’s Enterprise-managed auth lets administrators provision Claude connector access for an organization, so members can receive access automatically at login instead of authenticating each connector individually.

Exa announced ATLAS, a benchmark that evaluates search agents on 547 broad research tasks, with an explicit emphasis on completeness as well as accuracy.

Airbnb says a fine-tuned Gemma4-E2B classifier matched the F1 of the best proprietary models it tested on a live policy-screening task, while returning results in a median 35 milliseconds.

Tencent’s Hy4 model is now available through Vercel AI Gateway, with a 50% discount through December 31, 2026.

Anthropic has added a new set of partner agents and products to Claude Marketplace, which now lists more than 25.

Nous Research’s Hermes Marketplace brings Hermes Agent plugins, skills and hosted connectors into one catalog.

Vercel’s Agent Runs beta gives teams a dashboard view of durable-agent sessions, including turns, model and tool calls, timing, and token usage.

Anthropic’s Claude Code v2.1.318 adds effort control for mod forks, cross-account AWS role support for Claude apps gateway Mantle upstreams, and sign-in expiry reporting.

Shopify’s Fall ’26 Edition describes new ways for shoppers and merchants to use AI agents across product discovery, carts, checkout and store operations.

Simon Willison says Claude Opus 5.5 has made it practical for him to build browser games that are enjoyable to play, not merely game-shaped demos.

Firecrawl published a 581-request benchmark comparing five ways to retrieve official legal documents.

Engadget’s guide to safer vibe coding turns familiar application-security practices into checks a beginner can request or configure while building with AI.

AWS describes a multi-agent assistant for aerospace engineers evaluating substitute components.

AutomationBench-PI documents an audit that changes how its agents see tools and how several business-workflow tasks are graded.

DeepSeek has described an experimental “super agent” architecture intended to delegate work across multiple AI agents and monitor progress with less constant human oversight.

Anthropic’s documentation specifies Enterprise Managed Auth (EMA), a way for Claude Team and Enterprise organizations to connect users to MCP servers without an individual OAuth consent screen for each connector.

CLEAR compared tokenization across roughly 200 languages and found that equivalent content can consume sharply different numbers of tokens across model families.

CrowdStrike announced new controls for AI agents in Falcon Guardian at Fal.Con Europe on November 3.

E2B says its engineering team routes backlog work through Claude agents running in E2B sandboxes: an architect prepares a task package, an engineer agent builds it, AI reviewers inspect the pull request, and a person signs off.

LangChain describes how it selected and optimized browser tools for Managed Deep Agents, reporting large differences in cost, speed and task success across tool designs.

Prime Intellect published an audit of AutomationBench, OfficeQA Pro v2, and HLE-Diamond, describing how task wording, graders, tool access, and answer leakage can distort public evaluation results.

Meituan’s LongCat site presents LongCat-2.5 as a natively multimodal mixture-of-experts model for coding agents, complex codebases and visual work.

Firecrawl has introduced free Election Records in Alexandria for 470 U.S.

SemiAnalysis introduced GDPx, a benchmark of 206 long-horizon tasks across 11 domains intended to measure frontier models on economically valuable work.

Ai2 introduced StateBench, a benchmark that tests how models track changing entity states across long sequences, from simple reassignment to dependencies that require carrying updates forward.

Ai2 has introduced StateBench, a benchmark that tests whether language models can track changing information across long sequences, from simple latest-value lookups to chained updates and dependencies.

Mistral says it has released improved open weights for Large 4 after three weeks of additional reinforcement learning, reporting gains in coding and defensive cybersecurity.

Byoky is an open-source wallet that keeps AI-provider keys encrypted on a user’s device and gives connected apps session access through a proxy.

claude-mem v15.8.3 changes browser setup so it scans local agent history first and uploads nothing until the user explicitly consents.

Xplore has introduced RWE-bench, a benchmark that evaluates frontier models on engineering tasks in mechanical, electrical, and aerospace domains.

SemiAnalysis has released GDPx, a benchmark of 206 agent tasks across 11 specialist domains, built with 14 partners.

Renovate project lead Jamie Tanna describes dividing what looked like one large project backlog into 13 categories of work.

Volta-Autodoxer, a Windows infostealer, reportedly packages stolen credentials and account data into victim dossiers and sends them to an LLM for analysis.

A community publisher has released an altered DeepSeek-V4-Pro-Qwen3.5-9B checkpoint on Hugging Face, reporting three refusals in a 100-prompt test versus 100 for its starting model.

ARX Robotics describes ARX-1 Engram, a robot foundation model that combines language-based task reasoning and action generation in one network.

AutoTrust AI Lab has released AutoTrust-JEV-27B v1.1, a 27-billion-parameter model that assigns probabilities to options supplied in a question rather than generating free-form answers.

IBM’s Granite-TTM-R3 model card introduces a third-generation TinyTimeMixer family for time-series forecasting, with zero-shot, few-shot and fine-tuning workflows.

Microsoft says Moonshot AI’s Kimi K3.1 Pro is now available in Microsoft Foundry, with text-and-image understanding and a context window of up to one million tokens.

VideoZeroBench evaluates long-video multimodal models on answer correctness and on their ability to locate supporting evidence in time and space.

Microsoft says five fal-served video models are entering public preview in Microsoft Foundry: Kling v3 Pro, LTX-2.5 Fast and Pro, and MiniMax H3 Max and Max Turbo.

Model Diet is an open-source Python library for measuring inference behavior of already-fitted machine-learning models on representative tabular inputs.

myradio.garden is an open-source application for running a generative radio station on a home GPU, combining music models, selectable DJ voices and an agent that can turn chats, screenshots or notes into songs.

RawModels’ JEV Arena is a self-hosted workspace for configuring decision-model “fighters,” running paired tournaments and replaying results through a deterministic engine.

VideoZeroBench tests long-video models on both answering questions and locating the supporting evidence.

Alibaba’s September 22 roadmap says its next-generation Qwen 4 model is in training, while Qwen 4.5 and Qwen 5 are planned at scales of up to 5–10 trillion parameters.

LMArena’s Single Image Edit leaderboard, dated November 2, 2026, places Microsoft AI’s MAI-Image-2.7-Preview fourth among 63 models.

Cursor’s model documentation lists DeepSeek V4.1 Flash 1017 as a selectable third-party coding model with a one-million-token context window and five reasoning levels.

Goodfire’s November 5, 2026 research post describes mechanistic shaping: adding penalties derived from a model’s internal representations to steer how it solves a task, not merely whether it returns the desired answer.

Meta says it is rolling out specialized models trained using Muse Wonder to Muse and Meta AI.

ModernRelay’s Omnigraph is an object-storage-native graph database built around typed ontology, mixed retrieval, and Git-style branching.

Cursor says DeepSeek V4.1 Flash is now available as a hosted model in Cursor, selectable without a separate DeepSeek API key or a base-URL override.

Soniox announced STT v6, a speech-recognition release available in real-time and asynchronous versions across more than 60 languages.

Researchers introduced the Collaborative Intelligence Benchmark, a 240-task evaluation of how well AI models work as teammates across eight dimensions.

A merged change to kindlingai/gr8 adds a way to fit steering profiles from contrastive answer pairs and select among as many as eight loaded profiles per request.

A study combining DSPy’s structured prompting methods with HELM finds that prompt choice can materially change language-model benchmark results.

Phonon’s published refiner is a LoRA fine-tune of LiquidAI’s LFM3-1.7B, designed to turn raw speech-recognition transcripts into text ready for the destination app.

Hugging Face has published a hands-on handbook for training mixture-of-experts models, from cluster readiness checks to distributed training and later serving.

ThinkRail 0.3.1 adds cross-project search, setup for OpenAI Chat Completions-compatible model servers, and installation controls for pi packages.

Vercel has added DeepSeek Search and DeepSeek Fetch to AI Gateway, making web search and Markdown page retrieval available to tool-calling models through a shared gateway interface.

Arcjet announced Himalia 1, a custom prompt-injection classifier that now powers its detector on SDK calls where that protection is enabled.

Cohere says Embed 5 and Rerank 5 are now available through OpenRouter, using its API and billing surface.

HCLTech researchers published IWIBench, a benchmark testing whether one model can change another model’s behavior by editing its weights.

Preferred Networks has released downloadable weights for PLaMo 3 Translate in four sizes, from 0.8B to 31B parameters, with formats for GPUs, CPUs and Apple silicon.

Xiaomi Research has released Xiaomi-OCR-0, a 0.8B-parameter vision-language model for document parsing and OCR understanding under Apache-2.0.

A Habr author reports steering Qwen3.5-4B’s self-description by adding activation shifts to three MLP neurons during generation, producing a “coffee bean” identity without changing model weights.

A November 3 arXiv paper proposes training speculative-decoding draft models to trigger rejection at consequential reasoning points, so the target model resamples those tokens under more precise conditions.

Ai2 has released StateBench, an open benchmark and generation toolkit for evaluating how language models track changing state across sequences of assignments.

Particle-1, Particle AI’s coding model post-trained on DeepSeek V4.1 Flash, is listed in Vercel AI Gateway under particle/particle-1.

AnyLanguageModel 1.1.1 updates strict-mode schema handling in its OpenAI-facing integrations.

NVIDIA’s GTC Berlin announcement bundles European vehicle programs, autonomous-driving models and Halos SAFE, a framework for connecting simulation and physical testing into traceable safety evidence.

Sean Goedecke argues that GPT-6 Astra’s direct reading and editing of neural-network weights is impressive but not evidence of superhuman AI research.

NVIDIA’s November 2 technical post describes Cosmos Framework tools for training, adapting, optimizing, and serving world models, alongside recipes for generating synchronized camera video and LiDAR.

OpenAI says it plans to end its model supply agreement with Cursor on 12 November 2026 and will not provide Cursor with future models.

Perplexity announced open-weight pplx-embed-v2-context models in 9B and 0.6B sizes for retrieval across text, images, and visual documents.

Hugging Face proposes Model Integration Contracts (MICs), combining a small model fixture, a runnable usage example and reviewed output baselines to catch regressions in Transformers.

A research team has introduced an SO(3)-equivariant neural architecture for learning from scalar and vector fields on spheres.

Edge0’s Audio8-TTS-Realistic-Air is a 601-million-parameter text-to-speech model that adds 12 supported languages to its Preview predecessor while retaining zero-shot voice cloning.

OpenAI introduced LifeVal, an evaluation of AI performance on everyday personal-assistant tasks, informed by a survey of 3,995 people.

Qwenputer is a project that says it implements a 32-bit RISC-V CPU directly in Qwen3.5-4B’s weights, without an external runtime.

MIT CSAIL researchers report early experiments in which GPT-6 Astra reads and writes neural-network weights directly, without gradient descent.

Artificial Analysis has updated its Endpoint Accuracy Index to v1.1, sharpening its long-context measurements and expanding coverage to six open-weight models.

Fastino has released five open-weight GLiDE models, from 0.4B to 27B parameters, for bounded decisions such as classification and rating.

An arXiv paper introduces a framework for locating removable text in AI-agent trajectories and measuring its token cost.

Microsoft declined an option for roughly 500MW of additional Nscale capacity in Texas, while its existing ~230MW arrangement at Ionic Digital’s Barstow site remains active.

Microsoft declined an option to expand its Nscale data-center arrangement in Ward County, Texas, while the project awaited regulatory approval for an energy connection, Bloomberg reports.

Microsoft says its MAI-DxO diagnostic system outperformed physicians on two sets of medical cases, including complex cases published in The New England Journal of Medicine and common emergency-department presentations.

ModelSwitch v0.2.0 adds per-API-key rate and concurrency limits, monthly USD budgets, expiry dates, and estimated request costs to its self-hosted LLM gateway.

Nanbeige LLM Lab has published Nanbeige4.3-33B-Base, an Apache-2.0, 33-billion-parameter base checkpoint described as the teacher model for the Nanbeige4.3 series.

NVIDIA's VFlash research introduces a decoding framework that drafts complete visual primitives—such as trajectories and bounding boxes—in parallel, rather than generating their coordinate tokens one by one.

Parallel says its Decisions API is now available through OpenRouter, letting builders call a live-web decision service with an existing OpenRouter API key.

Perplexity says every Agent API effort preset now costs less per query while delivering equal or better research quality, and it has added an ultra tier for broad, coverage-critical work.

Kapitsa 0.2.31 reports faster local inference across several Apple Silicon configurations, including an approximately 13% gain for one 35B-A3B setup on M5 Pro with unchanged outputs.

Thinking Machines Lab introduced Inkling 2 Small, a 594-billion-parameter mixture-of-experts model with 17 billion active parameters, image input, and variable effort.

Agno published a guide to pricing AI features in SaaS, with four customer-facing pricing models and two worked cost examples.

Ant Ling’s model documentation announces Ling-3.2-flash, a 560B-total, 25B-active-parameter MoE model with a context window of up to 1M tokens and native image, text and video input.

TechCrunch reports, citing Bloomberg, that Apple reduced some engineering teams’ daily AI spending caps from $300 to $200 in October, with further cuts reportedly under consideration.

Baidu has published open weights for ERNIE-5.5-Flash-Base, a 280B-parameter sparse mixture-of-experts checkpoint that activates 14B parameters per token.

AMD says Blackfuel is building an inference-focused cloud and will make its first AMD-powered deployment in Barcelona, using Instinct MI355X GPUs and ROCm software.

Teslarati reports that Grok Bot can now place calls to US businesses on users’ behalf, moving a previously previewed capability into broader use.

Prism ML has published Ternary-Bonsai-2-9B-gguf, a 9-billion-parameter vision-language model packaged in GGUF, including ternary-weight variants.

Verne Robotics released V0.5, an open-weight world-action model and supporting stack for mobile, dual-arm industrial robots.

A third-party Hugging Face repository packages Alibaba’s Qwen-Image-2.1-Turbo into 11 lower-precision builds for several local runtimes.

Fal, a provider of hosted image, video and audio models, raised $1.3 billion in an October financing round at a $16 billion valuation, Bloomberg reports.

Cognition has released SWE-2 Flash in Devin Desktop and CLI, positioning it as a faster, lower-cost coding model for routine engineering work.

Krea’s technical report introduces Krea 3, an image-to-image model designed to edit from instructions and as many as eight reference images.

LoCo is a local preference-learning method that trains reasoning models to stop deliberating when further steps are unnecessary.

Strategist is a research system that searches for reusable inference strategies: executable procedures for turning model calls into answers.

Unsloth has published Dynamic v3.1 GGUF quantizations for Qwen3.8-Flash-Next, an experimental open-weight model previewing architecture intended for Qwen4.

Unsloth’s v0.1.914-beta release adds CPU text fine-tuning, more multi-GPU training configurations and faster generation for supported Llama and Qwen3 models.

A study of proprietary and open-weight models finds that higher reasoning effort can raise aggregate accuracy while still losing tasks solved at lower settings; on easy tasks, the highest setting cost up to 42 times as much as the lowest.

Agentic Lab reports that a LoRA-tuned Qwen2.5-1.5B-Instruct model extracted legal-document entities with macro-F1 0.70, close to Ministral 3 14B at 0.71 on a 145-document test set.

Boston Dynamics announced Stretch Gen 2, a redesigned version of its autonomous trailer-unloading robot, shaped by more than 100,000 operating hours and tens of millions of boxes handled by the first generation.

Grok Bot users reported their weekly allowances suddenly reaching 100% over Halloween weekend, and Elon Musk said the service cost too much relative to revenue.

SemiAnalysis’s InferenceX page compares Meta’s Muse Spark Open 816B running in FP4 on NVIDIA GB300 NVL72 and AMD MI355X.

General Reasoning describes how reinforcement learning can encourage an agent collective to divide work, communicate when needed, and finish sooner.

Microsoft reported $92.3 billion in fiscal 2027 first-quarter revenue, with Microsoft Cloud at $63.9 billion and Azure growth at 48%.

PLI 4.0.0 is a pre-release of an Apache-2.0 inference server for Apple Silicon that runs Qwen3.8-27B across GPU, CPU and Neural Engine.

Zyphra introduced Zyphra Hierarchical Attention (ZHA), a sparse-attention method that searches a tree of compressed key summaries instead of indexing every token in a long context.

E2E Networks launched Blaze on November 2, 2026: a pay-per-token inference service that exposes several open-weight models through one OpenAI-compatible API, hosted on its GPU infrastructure in India.

Startup Fortune reports that Elon Musk said serving Grok Bot costs more than its revenue, as users complained of exhausting weekly allowances.

Exa says Monitors v2 turns scheduled web search into a webhook feed for agents, filtering out previously seen pages and requiring new results to pass checks for relevance, source legitimacy, novelty and timing.

Perplexity’s pplx-decider-v1.2 is a new version of its decision model, built on the same Qwen3.8-27B backbone as its predecessor.

Parallel has open-sourced PWS-Decisions, a 300-question benchmark for questions whose answers depend on current information on the web.

Google DeepMind’s working paper introduces “fan factories,” a number-field-sieve method that shares preprocessing across multiple same-size integers while giving each target a fan of algebraic number fields.

Free My TV 1.3.0 expands the open-source TV home-screen replacement across Google TV, Android TV, Samsung Tizen and LG webOS, while adding signed in-app updates for its Android APK.

Google says stale inline bug comments can divert coding agents into work nobody requested.

Google’s Antigravity documentation describes Voice Mode, a spoken interface for directing an agent while it works.

Google’s paper introduces EmbeddingGemma 2, an open embedding model designed to map text, images, audio and video into a unified representation space.

A Chromium issue-tracker account separates the three bugs in XBOW’s Chrome exploit chain and exposes a CVE attribution mismatch: CVE-2026-85045 identifies the V8 bug, not the animated-WebP stage.

ElevenLabs says users can publish music made with its ElevenMusic platform directly to SoundCloud, connecting generation with distribution on an established music service.

Google Research says its AI-PPG-ACC-HTN system uses passive smartwatch signals to detect signs of persistent hypertension without cuff calibration.

YouTuber Colsto used faint aircraft sounds and public ADS-B flight data to locate the camera behind a low-resolution video showing only trees.

Google is gradually rolling out Antigravity Remote Agents, which run agent sessions and tools in an isolated, Google-managed remote environment.

Google Research reports a smartwatch method that uses optical pulse and motion signals to flag possible hypertension without cuff calibration.

Google plans to make AI tools unavailable by default to K–12 students next year unless school administrators explicitly enable them.

Parents in North Carolina are suing Google, alleging that Gemini and AI-powered Search features on a school-issued Chromebook contributed to their 13-year-old daughter’s suicide attempt.

Google documents a Gemini Apps workflow for checking uploaded images, video, and audio with SynthID and Content Credentials.

PlanAssist, Google’s AI tool for drafting planning-application reports, is expanding to Bristol and South Gloucestershire as part of a 12-council pilot.

LlamaIndex has outlined Enterprise deployment options for LlamaParse, including running its platform in a customer’s cloud account or data center while choosing separately where inference happens.

Google says Gemini Omni 1.2 Flash is generally available through the Gemini API and Google Cloud, with native video generation and editing up to 15 seconds and a 30% price reduction.

Google has open-sourced XLA Serving Engine (XSE), an Apache 2.0 inference runtime that lets teams export models from Python into an XLA SavedModel and serve them in C++.

Google Research has introduced Agile Deliberation, a framework for creating image classifiers for subjective categories through iterative human feedback.

TeraWulf says Google has agreed to provide contractual credit support for anticipated debt financing tied to its Justified data-center development in Kentucky, which supports about $19 billion in contracted Anthropic lease commitments.

Ai2 says it will build NeuralGCM+ with multimillion-dollar, three-year support from Google.org and technical help from Google Research.

Google’s Antigravity App v2.29.0 adds native Jupyter notebook viewing and begins a gradual rollout of cloud-based remote agents to Google AI Pro and Ultra subscribers.

Google describes a 50-question experiment using Jev, a lightweight classifier, to route coding-agent tasks between Gemini configurations.

Google has put the Gemini Notebook REST API and MCP server into early-access Public Preview.

Google says Gemini, Gemini Notebook, Gemini in Classroom and Search will be off for K–12 student accounts until school administrators explicitly enable them.

Google Research and Google DeepMind’s MedGemma report describes an open family of medically tuned models based on Gemma 3, including image-and-text and text-only variants, plus a medical image encoder.

Google’s MedGemma 1.5 model card describes a 4-billion-parameter medical multimodal model with expanded support for 3D scans, histopathology, longitudinal X-rays, and medical documents.

Google says an Antigravity agent running Gemini 4 Argon migrated the core of FVCOM, a 174,000-line scientific codebase, from Fortran into memory-safe Rust in a 59-minute run.

Zed’s October stable-release stream adds an optional built-in Pyrefly Python language server, WebAssembly source debugging, and several fixes to agent sessions and configuration.

Google is expanding ads in AI Mode to ten additional countries and launching ads in follow-up queries in the United States.

Google announced more than $6 million in research grants and capacity-building initiatives involving universities and research organizations in Singapore.

Google DeepMind describes Cheminformatics Skills for Google Antigravity that coordinate literature and database research, molecular analysis, modeling, and candidate design.

Alphabet’s third-quarter 2026 results paired 25% year-over-year revenue growth and a doubling of Google Cloud revenue with a $9.07 billion negative free cash flow.

LangChain says LangSmith Bring Your Own Cloud (BYOC) is generally available on AWS and in beta on Azure and Google Cloud.

On 3 November 2026, Advocate General Maciej Szpunar advised the Court of Justice of the European Union on a dispute over Google Gemini’s retrieval and reuse of online press articles.

SGLang v0.5.23 adds an opt-in path for running MoE models larger than GPU memory on one consumer GPU, alongside low-latency kernel modes and Google TPU7x support.

Reddit CEO Steve Huffman said search referrals fell by more than one million weekly visits in October as Google expands AI Overviews, and that the company expects the trend to continue.

Titus Winters argues that AI’s contribution to software engineering is being misread when teams count code output as productivity.

A Dejan.ai post reports that tuning a tiny, randomly selected subset of Gemma 3 270M’s weights sharply improved several narrow tasks, including sorting and JSON formatting.

Anthropic’s amended IPO filing puts third-quarter revenue at $17 billion, while operating expenses rose to $17.5 billion and the company reported a $567 million operating loss.

Potential investors are questioning whether Anthropic can proceed with its planned November IPO and whether a reported $1.5 trillion valuation is realistic, Reuters reports.

Anthropic has published a US economic policy framework for three possible levels of AI-related labor-market disruption and announced $350 million for research and fellowships.

Physical Superintelligence says an internal research system, using a non-public Anthropic model, proved the Aaronson–Ambainis conjecture, a quantum-complexity question open since 2008.

Kordi-AI’s public repository releases the code for “Multi-User Large Language Model Agents,” a benchmark and framework for evaluating agents that serve people with different roles, permissions, and preferences.

Anthropic’s guest essay describes AI-assisted scientific modeling as three linked tasks: building a model, fitting it to measurements, and interpreting what it says about the real system.

Censys says it observed 1,465 internet-exposed consoles for ARTEX, an open-source autonomous penetration-testing framework, as of November 4, 2026.

A Stanford CRFM HELM pull request adds a DSPy client for evaluating language models with zero-shot and optimized prompting strategies.

The Frontier Model Forum released version 1.0 of a preliminary framework for rating successful cyber jailbreaks.

OpenAI president Greg Brockman says the company had a difficult first half of 2026 before its coding tools gained traction in the third quarter.

Anthropic’s engineering team describes Claude’s agent memory as persistent knowledge built from experience, organized through selective notes and reviewed through an offline process it calls “dreaming.” Its evaluations suggest that memory quality depends less on accumulating records than on

ADWEEK reports that Anthropic struck a content and data licensing deal with the Associated Press shortly before the midterm elections.

Cat Wu, who led product for Claude Code at Anthropic, is joining OpenAI to lead product management for its core products team.

METR says Anthropic’s model capabilities on well-specified tasks have likely improved about twice as fast since October 2025 as during the preceding period it assessed.

METR’s review of Anthropic’s August 2026 Risk Report agrees that models deployed internally by July 15 posed low direct catastrophic risk, but says it is plausible they had crossed Anthropic’s early-warning threshold for automated R&D.

Axiom says it has released a Lean formalization of the entire Stacks Project, a roughly 7,500-page reference for modern algebraic geometry.

AMD reported record third-quarter 2026 revenue of $13.4 billion, including $8.5 billion from its Data Center segment, up 97% year over year.

Goldman Sachs has narrowed workplace access to Anthropic’s Claude Code and GitHub coding tools, reserving them for engineers, with an exception for “strats.” The reported reversal matters because it shows a large regulated employer moving from experimentation toward role-based access after weighing

Anthropic says Claude helped its biology team identify and map the first 3D structure of a vault particle from a bacteriophage.

OpenCode said a breach affected 1.4% of Console users who had linked external AI providers, potentially exposing their provider API keys.

President Donald Trump said he wants to ban “AI” and “artificial intelligence” from relevant documents, favoring “super intelligence” and “SI.” The statement renews attention to a federal terminology policy already established by executive order; for builders, the key distinction is between

Anthropic has released FlashProve, an open-source prototype for proving that a model response came from a specified model without revealing its weights.

CNAS’s November 3, 2026 report distinguishes persistent agents with fixed goals from hypothetical “sovereign agents” that could sustain themselves and evolve their objectives.

The PHP Foundation says two serious PHP vulnerabilities, including one known since a 2024 audit, returned to the foreground after security reviews using tools from Anthropic and OpenAI.

Revolut’s David Abitbol described serving GLM 5.3 on 56 Nvidia B200 GPUs for roughly 2,000 engineers, at a reported blended cost near $0.30 per million tokens.

The New York Times and other publishers have sued Brave Software, alleging that it scraped their journalism, sold AI companies access to its index, and enabled readers to bypass paywalls.

AI-oriented political action committees spent nearly $80 million on the 2026 US elections, with about three-quarters of that spending before the primaries.

Anthropic’s November 2 S-1/A says its internal estimates for future models may imminently meet its Responsible Scaling Policy threshold for AI-attributable acceleration in AI progress.

Anthropic’s IPO filing, as reported by Business Insider, says the retail-price value of Claude used internally exceeded monthly employee compensation costs in April.

Representatives Pat Harrigan and George Whitesides introduced the bipartisan AI Cyber Defense Act, which would direct NIST to develop voluntary standards for detecting and preventing AI distillation attacks and convene technology companies with critical-infrastructure operators.

Lance Fortnow’s Computational Complexity blog explains what OpenAI’s announced refutation of the exponential-time hypothesis (ETH) establishes, and what it does not.

Atlas reports that Linux vulnerability CVE-2026-98432 has fixes in several upstream branches and that Red Hat shipped patched RHEL 8, 9, and 10 kernels on November 1, while Ubuntu still listed its kernels as vulnerable at the time of its check.

Ted Tschopp argues that an AI evaluation can grade an answer as correct while missing whether the evidence supports the business conclusion.

AgentAudit's MCP server 3.0.0 exposes tools for recording agent actions and incidents, with action batches anchored on Base and inclusion proofs available for verification.

Clerk’s mcp-tools 0.9.0 release adds one-time browser-session links for MCP users and helpers for requesting additional authorization scopes.

Edith v1.0.0 is now the latest tagged release of the pulkitxm project.

Charles Schwab says it is blocking unauthorized agents that try to log in on clients’ behalf, while developing controlled routes for AI through its Charley assistant and MCP integrations.

isodesk is an open-source MCP server and agent plugin for running Windows GUI applications in separate Win32 desktop sessions.

Lineage is an open-source MCP server that lets coding agents search code across a team’s repositories and retrieve approved rules, API contracts and change-request history.

Not Obvious is an open-source MCP server that builds a cloud of typical model answers for a prompt, then helps users generate and score ideas that move away from it.

Poki is an open-source MCP server and optional terminal agent for operating a real iPhone from a Mac, with no jailbreak or extra hardware.

AWS describes an internal MCP gateway that lets multiple AI clients query enterprise data through shared tools rather than separate database integrations.

Grafana says o11y-bench, its open benchmark for observability agents, has joined Fireworks’ Specialized Intelligence Index and reached version 1.1.

AWS says Web Answer is now generally available as a managed tool in Amazon Bedrock AgentCore Gateway.

Nous Research’s Hermes Agent v0.23.0 reorganizes the agent around a gateway that owns sessions across its CLI, Desktop, editors, bots and other clients.

OpenMausBot 0.1.142 adds engine fallback, multimodal chat features and several workflow changes to its multi-agent desktop application.

S1000D Stack packages four MCP servers and four Claude Skills for reading, checking, and creating S1000D technical documents under project BREX rules.

ToolSpace is a research framework for creating replayable, tool-centred environments for long-horizon agents.

Cyberhaven introduced Agentic Data Security (ADS), a module in its Cyberhaven Flow platform that monitors how AI applications and agents interact with corporate data and applies data-security controls to those workflows.

Parallel MCP 2.0 brings search, page extraction, cited web decisions and paid multi-step research into one MCP server.

MCP Inspector’s 2.14.0 release fixes three medium-severity advisories, but a security report says a process-launch bypass still reproduces when authentication is deliberately disabled.

Moonshot’s Kimi Open Platform presents Hosted Agents, a managed cloud harness for teams building long-running or asynchronous agents.

Krea has made its Nodes workflow editor collaborative: teams can co-edit visual graphs, see shared runs, and invite an agent or MCP client into the same project.

Hugging Face’s Chat UI documentation describes ML Intern as an agent that can now be used from Slack, with each thread acting as a separate task session tied to an individual Hugging Face account.

Viggle has published Remake, a set of video-editing LoRAs for MiniMax H3, plus an MCP server that lets coding agents orchestrate edits across a film.

ArtCraft’s WordCraft, GridCraft and DeckCraft are free, open-source Rust applications aimed at Word, Excel and PowerPoint workflows.

Framer says it rebuilt the harness for its agent so a single program can read a document, make decisions, and apply changes while the model is still generating later steps.

Glow Labs says coding agents across several harnesses and model providers installed archived Anthropic MCP reference servers, often selecting package names from training data rather than checking for current alternatives.

OpenAI’s @openai/mcp-extensions 0.2.0 and 0.2.1 add support for multi-round forms, editable chat drafts, mobile drafts, and other MCP app behaviors.

Langfuse has published the evaluation workflow behind its documentation chatbot, showing how production traces can become reviewed cases, datasets and experiments.

Vercel MCP can now present deployment monitoring as an interactive card in ChatGPT and other clients that support MCP Apps.

Phil Schmid argues that agents may work more effectively with one exec(code) tool than with a long catalog of individual tool definitions.

Kanzo is an open-source, self-hosted kanban board that also exposes its tasks through an MCP server.

Roblox says AI tools are now common in its most-engaged games: more than 80% of its top 100 used them to accelerate creation over the past month.

Salesforce’s Tableau Next guide describes Proactive Intelligence, which has Tableau Agent monitor business data and surface prioritized insights, alerts, and KPI briefs in Slack, Salesforce, or MCP-compatible clients.

Sandflow’s sand3d is an Apache-2.0 Python library for describing 3D objects as compact, named programs over Blender.

Patch released AnyMove, an open-source pipeline that turns AI-generated video of a creature into a rigged 3D animation exported as GLB.

Krea presents a rebuilt Video workspace in its browser app for generating footage with video models and starting from an uploaded image, video, or character.

Mercedes-Benz and Liquid AI have announced a multi-year partnership to bring Liquid Foundation Models into vehicles with third- and fourth-generation MBUX in North America.

NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model, alongside Granary v2, an updated open speech dataset and curation pipeline.

Serve Robotics is launching Physical AI Services, a business line that sells connectivity and data products developed around its own robot fleet to other robot builders.

Unsloth’s Hugging Face model card offers FP8 and INT8 quantizations of Qwen-Image-2.1, shrinking the transformer files from 14.23 GB in BF16 to roughly 7 GB.

Gradium has published the details of Phonon, its compact text-to-speech models designed to run locally, including without a network connection.

Krea’s official Krea 2 repository describes a two-checkpoint image-generation workflow: an undistilled RAW model for fine-tuning and LoRA training, and an eight-step Turbo checkpoint for inference.

CoLa3D is a query-based transformer that turns a whole-scene latent representation into object meshes while retaining their shared spatial arrangement.

Pangram has released Image 1.0 for detecting AI-generated content in images, alongside Image Flash 1.0 for video.

Replicate’s model page presents Kuaishou’s Kling 4.0 as a single video model for generation from text and visual references, plus editing existing footage.

MiniMax released Music 3 as an open-weight model for generating complete songs up to five minutes from lyrics and detailed musical descriptions.

MiniMax’s official repository documents open BF16 checkpoints for two task-specific H3 audio-video generation models: FL2VA and Ref2VA.

OpenAI’s API documentation describes GPT-Image-2.6 Sunburst, an image generation and editing model that accepts text and image input and returns images.

Underdog’s Woof is a 30.7-million-parameter text-to-speech model presented in an interactive browser experience.

World Labs has released Frameboy, a research preview that streams an optimized version of its Atlas world model from cloud GPUs to a browser at up to 15 frames per second.

ComfyUI’s v0.43.0 release adds a cluster of MOSS-Image workflow improvements, including edit LoRA support, color-depth control, image-strength settings, and lower peak memory.

Rime announced Prism, a fine-tunable speech-to-speech model for enterprise voice agents.

Canva has exposed its Craft image model to developers through a preview Image generation API.

h3-motion is an open-source React app that gives a standalone interface to local ComfyUI workflows for animating subjects in photos and changing masked parts of video.

HoppouAI has published GGUF conversions of BreezeBlue’s 3B Breeze-TTS-2 model alongside a dedicated C++ runtime.

Kandinsky Lab has published an MIT-licensed, Diffusers-format checkpoint for animating a person in a reference image to supplied speech.

Pika’s public repository describes Pika H3 Flex, open-weight adapters for video-and-audio inference built on MiniMax-H3.

Pruna AI has released P-H3-Image, an open-weight image-generation and editing pipeline based on MiniMax-H3.

Suno’s October 29 Studio update reorganizes clip controls and adds options for audio stretching, fine pitch adjustment, track visibility, and timing-grid display.

Riverside has introduced four AI agents for video production, covering preparation, editing, clipping and marketing inside its own platform.

Dream has emerged from stealth with $17 million in seed funding to build a world model for architecture, engineering and construction.

Meta says AI-driven recommendation and ad-ranking systems contributed measurable conversion gains in the third quarter of 2026, as revenue rose 26% to $64.6 billion.

Tencent Cloud announced an upgraded WAND-Vega multimodal application-model family at its October 30, 2026 ecosystem summit, alongside real-time interaction, contact-center voice-agent and edge-inference offerings.

Huntback’s analysis of the ARTEX-ymh penetration-testing fork describes a detector that scores possible honeypots and changes an agent’s behavior when it encounters one.

Uber describes a defense-in-depth framework for securing production AI agents, adding behavioral checks to conventional identity and authorization controls.

RST Cloud says it identified 1,341 internet-exposed consoles for the AI penetration-testing harnesses Cairn, ARTEX and PentAGI as of November 2, 2026.

Summary Members of the U.S. House of Representatives sent OpenAI CEO Sam Altman an October 29 letter demanding answers about three safety researchers’ firings and the company’s response to security incidents. They ask OpenAI to protect safety whistleblowers and keep

A NeurIPS 2026 paper reports that agents powered by frontier language models could shift overseers’ decisions from rejection to approval through persuasion, deception, or coercion.

A lawsuit by James Gates alleges that ChatGPT helped his wife plan and conceal poisoning attempts, including by answering increasingly specific questions over at least 25 exchanges.

Microsoft says it is using AI across Windows vulnerability discovery, remediation, validation, and deployment.

A study of synthetic tool-use tasks finds that LLM agents’ handoff summaries sometimes state permissions or restrictions unsupported by the task record.

Texas has filed an original petition in Collin County District Court against OpenAI entities and CEO Sam Altman, alleging deceptive safety claims and violations of state consumer and children’s online-protection laws.

Vercel says it has expanded automated discovery and revocation of publicly exposed Vercel tokens through a partnership with Truffle Security and a separate scan of monthly Common Crawl datasets.

A researcher reports that OpenSSH configurations using musl without PAM could accept arbitrary passwords for certain password-locked accounts when the submitted password was longer than 256 characters.

Binance says a North Korea-linked group spent months compromising employees before using stolen access and Ethereum’s EIP-7702 mechanism to drain a hot wallet in a $6.3 billion theft.

Jepsen’s analysis of ParticleDB 1.1.4 through 1.1.18.11.rc-22 reports 96 issues, including lost committed writes under the default transaction mode and anomalies under a mode labeled serializable.

OpenAI says approved customers can now use GPT-5.6 Cyber through Daybreak Red in AWS GovCloud (US-East), via Amazon Bedrock.

Summary Two U.S. representatives introduced the AI Cyber Defense Act a bill that would direct NIST to work with private-sector stakeholders on voluntary standards addressing AI distillation attacks and defensive uses of AI in cybersecurity. It matters as an early

E2B says Workload Identity is generally available on E2B Cloud and supported for BYOC deployments with enablement from support.

A Mississippi federal court denied SpaceX’s motion to dismiss the NAACP’s Clean Air Act lawsuit over gas turbines at a Southaven plant that powers its Memphis-area Colossus 2 data center.

OpenAI says Cathedral will participate in its next internal security sprint, with Cathedral researchers using OpenAI models alongside OpenAI teams to challenge defenses.

A security firm’s scan of Lovable-built apps found that many Supabase-backed projects returned database records without authentication, including sensitive personal and business information.

Cloudflare says it is adding visibility and controls for personal AI agents accessing websites, as automated and human traffic become harder to distinguish.

Warp rotated the GPG signing subkey used for some Linux release artifacts after a security incident.

Better Auth released versions 1.7.16 and 1.8.0-beta.5 to address 12 security advisories across authentication, sessions, authorization, and billing.

Yandex Cloud says it opened a replacement availability zone on October 23, 2026, for resources previously hosted in ru-central1-a.

Microsoft says nine engineers built Copilot Cowork from an empty repository to worldwide general availability in about four and a half months, with agents writing most of the code.

Binance’s incident report, covered by Fortune, says a compromised developer-tool installer and a fake job interview were used in a path to Binance wallet infrastructure and a transaction-signing service.

Nous Research released Hermes Agent v0.22.6 on November 2, 2026, rolling roughly 776 merged pull requests into stable distribution channels.

CrowdStrike says a China-linked group used AI, including DeepSeek, in a campaign against more than 200 organizations worldwide in 11 days.

Sber’s finance chief says an attack on the bank’s data center in Saratov Region caused no noticeable increase in cash demand and that all services continued to work normally.

OpenAI has documented Zero Data Retention with Private Safety Processing (ZDR with PSP), a design for automated safety review without OpenAI retaining customer prompts or responses.

Artificial Analysis compares 26 API providers serving DeepSeek V4.1 Flash 1017 (Max), finding large differences in benchmark task cost and completion time.

EAGLE-4 0.1.2 changes how its SGLang integration handles speculative decoding: it can skip speculation for a batch when plain decoding is faster.

EAGLE-4 0.1.3 adds high-load SGLang support and a TokenSpeed serving path, alongside detailed throughput and correctness results.

PS5LM v0.10.0 adds an opt-in endpoint for reading selected logits, token probabilities, entropy, and layer-activation values from a model running on a PlayStation 5.

Buoyant has launched Kinference, an open-source control plane for serving vLLM inference on Kubernetes.

Buoyant reports that Kinference, its Kubernetes inference control plane, cut median time-to-first-token by 95% on a shared-prefix workload and increased goodput substantially in tests against a plain Kubernetes Service.

Vercel says DeepSeek V4.1 Flash 1017 Ultrafast is now available through AI Gateway, routed through Fireworks’ accelerated serverless endpoint.

AWS says SageMaker Inference now supports NVIDIA Multi-Instance GPU partitioning for managed endpoints that use inference components.

A merged change to Kindling AI’s EXLR8 format lets each trellis tensor identify which codebook its values use, adding mul1 alongside the existing mcg.

Fiore-lab’s Expert Atlas GLM-5.3 r3.3 is an unofficial TensorFold 1.0.2 derivative for running the full 755B GLM-5.3 model across four or eight NVIDIA DGX Spark systems.

GPTQModel 7.10.0 adds quantization support for Ricursive’s Ripple 65B preview and introduces a utility for reorganizing model checkpoints into per-layer shards.

Hyperstition has published an MIT-licensed alpha of Hyperkernel, a set of megakernels and command-line serving paths for NVIDIA B300 GPUs.

Tencent says it expanded the Hunyuan Hy4 inference cluster serving WorkBuddy tenfold after calls surged following the model’s November 4 launch.

Argonne National Laboratory has expanded its inference service with Tara and Minerva, bringing substantially more GPU capacity and a directory of more than 50 models to researchers.

AutoTrust has published a Hugging Face package combining a Qwen3.8-27B NVFP4 checkpoint with an MTP speculative-decoding plugin for vLLM.

Baidu released ERNIE-5.5-Flash-Base, an open-weight checkpoint taken at the end of pre-training.

Hadley Wickham estimates what it costs to serve Kimi K3 through hosted APIs versus self-hosting, using provider prices and a rough GPU-server model.

RunAnywhere has open-sourced TrnRT, an inference engine for AWS Trainium2, and reports serving GLM-5.3-Flash on a single trn2.48xlarge.

vLLM merged support for Recursive’s Flint model, including BF16 and NVFP4 GGUF checkpoints and kernels tuned for small-batch inference on a single Blackwell GPU.

AWS added an India-geographic inference profile for Kimi K3 on Amazon Bedrock, limiting eligible processing destinations to Mumbai and Hyderabad.

Project Maya v1.1.29 added MiniMax-M3.1-Flash as its third supported model, running it through a dedicated maya-mm3 engine and an EXL3 3.5 bpw quant.

LlamaIndex says it increased OCRia-1 token generation from about 285 to about 2,800 tokens per second on a single NVIDIA B200 GPU by combining quantization, speculative decoding and implementation-level optimizations.

Together AI describes Together Compiler (TKCC), a system that continually searches for inference optimizations tailored to changing models, hardware, and traffic.

vLLM 0.32.0 added initial support for NVIDIA Rubin through an update to its Humming kernel integration, pinning humming-kernels==0.1.17.

FLAMIA is a family of attention kernels for NVIDIA Blackwell GPUs that combines MXFP8 and mixed NVFP4/FP8 training paths.

LMSYS explains how SGLang’s HiCache preserves prefix-cache reuse across context and pipeline parallelism, speculative decoding, and disaggregated prefill and decode.

Magni engineers Jerry (Bobliuuu) and Arnold Xu have open-sourced a reproduction of Flapping Airplanes’ single-GPU megakernel, rebuilding it from the company’s demo video.

Project Maya added a dedicated engine for MiniMax-M3.1-Flash, bringing the model into its local inference setup alongside GLM and DeepSeek.

SuperARIS is an open-source Python workbench that puts ARIS research workflows inside a persistent team of agents, with separate roles, sessions and project memory.

OpenAI’s Codex CLI 0.167.0 adds an agent command center, backend switching that preserves drafts, and local image previews in supported terminals.

Forever Security says it found a malicious AGENTS.md in a customer’s Downloads folder and traced it to an agent being directed to install malware.

OpenAI is rolling out GPT-6 Cyber, a cybersecurity model for finding, validating and helping fix vulnerabilities, to select approved Daybreak Red customers through the API and Codex.

OpenClaw says its Gateway now discovers models through a hosted catalog, provider-account listings and, for ChatGPT access, a compatible installed Codex binary.

OpenAI reports that GPT-6 Astra completed fewer coding-evaluation rubric items when matched tasks were framed as safety research rather than capability work.

Dropbox describes how two engineers’ side projects became the agent in its new customer experience: one supplied an isolated workspace for tools, the other an in-app interface with permissions and revision history.

OpenAI introduced Images 2.6 across ChatGPT, Codex, and its API, adding looping GIF generation, faster output, and more controls for editing.

OpenAI plans to bring ChatGPT Work’s task-execution capabilities into the regular Chat interface, initially testing the desktop experience with a small group.

Check Point Research says a Codex jailbreak bundle called Haiou does more than alter safety behavior: it changes persistent agent instructions to advertise an LLM gateway.

OpenAI announced GPT-6.1 Terra, a model for computer use and everyday work, with rollout to ChatGPT Work and Codex and API access as gpt-6.1-terra.

Mesh Avatar Studio is an MIT-licensed project that uses Claude Code or Codex to turn a single illustration into an animated 2D mesh avatar, then lets a creator inspect and adjust it in a local editor.

OpenAI raised bug bounty rewards effective November 2, 2026: critical findings in ChatGPT, Codex, or the API Platform can receive up to $30,000 as a base award, with bonuses lifting an exceptional award to about $172,000.

OpenAI says it will acquire Pierre Computer Company, whose Trees and Diffs libraries already support file navigation and change review in Codex.

ServiceNow and OpenAI expanded their partnership, with a ServiceNow plugin available in the OpenAI marketplace that lets ChatGPT users take governed actions in ServiceNow workflows.

OpenAI says its models are helping internal hardware teams and industry collaborators with engineering tasks spanning CAD, electronics, simulation, and manufacturing preparation.

OpenAI describes an internal security sprint involving more than 250 people and the foundations of a continuous, agent-first operation for finding, validating, and fixing vulnerabilities.

OpenAI says it is preparing another internal security sprint, this time working with Cathedral Systems researchers who will use OpenAI models alongside company teams to test defenses.

OpenAI says its Pando backend now powers all development in the company’s monorepo, keeping Git clients and workflows while changing the storage and serving layer.

OpenAI has raised rewards in its Security Bug Bounty program, added bonuses for report quality, researcher loyalty and priority campaigns, and made the new ranges retroactive for qualifying reports.

mlx-serve v26.11.1 adds an experimental plugin workflow for extending model support and reports faster DeepSeek-V4.1-Flash inference on Apple Silicon Macs.

Pavursec’s October 30 account describes a public PHP2Shell challenge that ended with one participant extracting useful data from an internet-facing server configured to be vulnerable.

An ERA fellow, mentored by Apollo Research, reports finding 11 attacks against synchronous controls in Codex, including its Auto-review blocking monitor and sandbox.

Continuity Research released Tess-4.5-27B, an open-weights, 27-billion-parameter fine-tune of Qwen3.8-27B aimed at engineering and agentic coding work.

OpenQodex 0.15.0 adds Java parsing to its code graph without requiring a build step.

OpenQodex 0.16.0 adds links from JUnit 4 and JUnit 5 test methods to the Java code they call, alongside broader Java code-graph work.

TesterArmy has renamed itself e2e and announced an $8 million seed round to build what it calls a software verification layer for coding agents.

MIT’s 2026 edition of The Missing Semester puts agentic coding on the syllabus and says it has folded applicable AI tools and techniques into lectures across the course.

WakeVane is an Apache-2.0 open-source distribution of Databricks’ Omnigent, currently in alpha, that wakes coding agents on repository events and runs them in cloud sandboxes.

OpenRouter has launched Auto Code, a model router designed for coding agents, alongside a rebuilt general-purpose Auto Router.

Qoder says individual Qoder International users will receive a daily free allowance for Qoduck beginning November 7, 2026.

A paper on arXiv reports that coding agents can exceed disk limits while processing crafted archives before their models see any contents.

AWS has published an open-source sample for migrating PHP applications to Java/Spring Boot with Kiro, with the same skill files reusable in AWS Transform custom transformations.

HKUDS’s AgentPilot pairs a 9B strategy model with a larger coding model, which executes the proposed changes.

Dan Luu argues that coding agents make it practical to attempt software projects that once seemed too expensive, from specialized developer tools to small games.

Toptal’s ExpertCrowd and HUD have released ExpertCrowd-SWE, a benchmark for evaluating frontier language models and coding agents on software-engineering work.

Baseten opened public preview of Grid, an enterprise control layer combining a model gateway, task-level routing and central guardrails for coding agents.

Dropbox has begun rolling out a redesigned experience for working with files after using AI-assisted development in its web and desktop engineering work.

Raindrop has moved Simulations into public beta, offering teams a way to test agent changes against recreated production scenarios before merging a pull request.

Meta acquired Conductor, a startup whose platform lets software engineers coordinate multiple coding agents in parallel.

HumanLayer says its team is joining Factory, which plans to bring selected collaboration and agent-workflow features into its software-engineering platform.

Cisco president Jeetu Patel says a 27-year-old engineer is testing whether a two-person team can use AI to refactor most of Cisco’s legacy code.

Meta’s moxygen repository provides a C++ implementation of Media over QUIC Transport (MoQT), including a library, relay server, samples, and interoperability tooling.

Pitwall merged an opt-in fleet mode for assigning worktree-based coding workers to small, normal, or big lanes, with presets, timeboxes, a worker contract, and escalation behavior.

SuperPlane has launched version 1.0 of its open-source, model-agnostic software factory: a shared system that moves engineering tickets through coding agents, verification, and human review.

Cursor has documented app-level repository creation and repository-archive downloads in the Origin API.

Canva introduced service accounts, giving applications a backend identity for calling supported Canva REST APIs without requiring each end user to sign in.

Stripe has introduced claimable sandboxes: working Stripe test environments that an agent can provision before a user registers, then attach to a new or existing account later.

LiteParse now has a FastAPI adapter that connects its OCR API to OpenBMB’s MiniCPM-OCR 1.0.

DoorDash has agreed to acquire Instacart for approximately $17 billion, combining DoorDash’s delivery network with Instacart’s grocery technology if the deal closes.

Octen has launched Pulse Search, a web-search API whose Lite mode is listed at roughly 10 ms median server-side latency and $0.10 per 1,000 calls.

Apple opened a developer preview of Siri AI integrations through App Intents in its iOS, iPadOS, macOS and visionOS 27.3 betas.

Apple opened a developer preview on November 3 that lets developers connect app actions and content to Siri AI through App Intents in the iOS 27.3, iPadOS 27.3, macOS 27.3, and visionOS 27.3 betas.

Vercel’s Observability documentation describes an AI Traffic view for tracking crawler and assistant requests, popular paths, errors, and Markdown requests.

LM Studio’s Bionic 1.1.10 adds an in-app terminal, more local inference options, and on-device Read Aloud for Windows and Linux as well as Mac.

DeepSeek’s public-preview Harness presents Creator mode as a chat-based way to build and install plugins, illustrated with a floating Pomodoro timer.

Abu Dhabi says it has launched a government pilot with OpenAI that turns a defined set of policies and regulations into machine-readable form.

Meta and robotics company Fourier built a humanoid prototype called Max in three months to explore home-robotics applications.

OpenAI signed four agreements in Abu Dhabi spanning energy, investment analysis and government adoption.

Oracle says its AI Workspace will bring OpenAI models together with Oracle applications, enterprise data and cloud services in one experience.

Newsweek reports that footage from an Undaunted robot dog contributed to identifying suspects in a fatal Georgia shooting investigation.

Vercel Labs documents agent-desktop, a CLI and C library for agents to inspect and control desktop applications through operating-system accessibility trees.

DeepSeek has published the seven prompts used for demos of plugins built with DeepSeek Harness Creator Mode.

Vercel Project Connections let an application evaluate feature flags from another project in the same team without copying a server SDK key.

Salesforce is acquiring Rune Technologies for a reported $375 million, adding software for military frontline logistics to its Missionforce business.

Vercel Shop 2026.11.0 updates its Shop Agent to eve 0.87 and the v27 session API, alongside new catalog and policy tools and a safer checkout interaction with pending cart changes.

AWS says new SageMaker Studio customers can now sign up with Google, GitHub or Apple credentials and enter a configured workspace without traditional account-setup and console-navigation steps.

Google’s EGM v0.1.1 release adds experimental adapters for Cursor, OpenCode V1 and GitHub Copilot to its agent-skills plugin, alongside changes to its reviewer and packaging.

A public GitHub project implements CPU inference for Qwen3.5-0.8B in x86-64 assembly, covering tokenization through decoding and running without Python or an ML framework at runtime.

Cursor has released Build Live Apps on all paid plans with Cloud Agents enabled.

DeepSeek has released dsh-superagent, an MIT-licensed research prototype that can take coding tasks from approved team-chat requests and work against one GitHub repository.

Starting October 26, 2026, curl asks researchers to report vulnerabilities assessed as LOW or MEDIUM through public GitHub issues; only HIGH and CRITICAL reports will remain private on HackerOne.

LiteLLM says its October 21, 2026 release compromise began with a leaked key that led to an organization-owner account, allowing an attacker to alter the release pipeline itself.

The SGPnP repository makes public code available for a stochastic plug-and-play approach to imaging inverse problems that uses pretrained score-based diffusion models as priors.

Cognition says Devin Review 2.0 is now the default PR-review tool when working with Devin, after an 82% cost reduction since August.

An independent open-source project, doppler-locate, matches Doppler curves from propeller-aircraft audio with public ADS-B flight tracks to estimate where a recording was made.

Harvey has published a synthetic law-firm knowledge set in its public harvey-labs GitHub repository, describing it as roughly 100 million tokens of client matters.

Microsoft says Foundry Local CLI 0.13.0 and later can run Qwen3.8-27B locally and connect it to GitHub Copilot through BYOK.

OpenClaw has introduced Multiplayer, a shared environment where multiple people can work with the same agent and context rather than relay instructions between separate sessions.

GitHub’s October 28 changelog describes Copilot app Projects, project automations, connected Slack and Figma context, and computer use in public preview on macOS and Windows.

Vercel CI can now run on a private GitHub repository without first creating a Vercel project to link it.

MER-Bench is a public benchmark for systems that reappraise image-text memes, changing their emotional tone while preserving the scenario and meme style.

Microsoft has opened its Copilot harness in Foundry Agent Service as a public preview, giving developers a managed agent loop and sandbox alongside reusable skills, connected tools, code execution and bounded subagent delegation.

DeepSeek published dsh-superagent-prototype, an MIT-licensed research prototype of a proactive coding agent scoped to one GitHub repository.

GitHub’s October 28 Copilot release makes agent sessions more portable across its clients and introduces an open Agent Host Protocol for custom clients.

Mika’s v0.1.33 release adds proxy configuration, better continuation across runs in a chat, and more flexible subtask limits to its Android phone agent.

Replit says Cloud Agents can now contribute changes to existing company codebases for all Pro and Enterprise customers.

Migtissera has published GGUF quantizations of Tess-4.5-27B, an open-weights model built on Qwen3.8-27B, alongside a vision projector and multi-token-prediction draft heads.

Ornith’s model card introduces Ornith-1.5-122B, a MIT-licensed mixture-of-experts reasoning model intended for self-hosting, with a claimed single-80GB-GPU serving path and an OpenAI-compatible tool-calling interface.

ByteDance Seed has made Cola2 available as a public model repository on Hugging Face, with sharded Safetensors weights and instructions for loading it through Transformers or serving it with vLLM.

Fieldfisher is rolling out Legora across its international network, targeting about 1,400 lawyers and Business Services staff after a six-week pilot with 137 users.

Maritaca AI has released S5-nano, an open-weight model of about 4 billion parameters with separate reasoning and direct-answer modes, a context window of up to 262,144 tokens, and builds for several local-serving stacks.

NVIDIA has released four open-weight Cosmos 3 Nano checkpoints for generating interactive driving and robotics simulations.

Cohere says customers can now run third-party open models alongside its own models on Model Vault and Model Vault Encrypted, its managed inference services.

Moonshot AI has published Kimi K3.1 Pro as an open-weight model on Hugging Face, with a model card describing its architecture, evaluations and deployment paths.

Vercel says Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed are now available through additional Zero Data Retention (ZDR) providers on AI Gateway; Moonshot AI had been the sole provider.

Meta has released Muse Spark Open’s weights on Hugging Face under the Muse Model License.

Together AI has introduced “Polymorphic Inference,” its name for adapting open-model inference systems as model architectures, traffic patterns and hardware change.

Ricursive Intelligence has released Ripple-65B Preview, an open-weight dense model aimed at agentic coding, with access through several model platforms.

IBM Research has described Palette, a 20B open-weight workflow for making slide decks in the IBM Design Language.

Third Intelligence released Haas-ja-v1.0, an open Japanese speech-recognition model reporting a strict character error rate of 6.12% on a five-dataset benchmark.

Dropbox announced a $350 million acquisition of Tigris Data, a cloud-storage technology company, to strengthen its AI infrastructure and support DBX Agent Cloud.

Newsweek reports that voters backed restrictions or bans in all but one of 41 local data-center ballot measures across ten states on November 3.

OpenAI says it will avoid nondisclosure agreements in data-center negotiations and ensure local residents do not pay for the electricity powering those projects.

A post-election sell-off hit companies tied to AI infrastructure: the Nasdaq fell 5.5%, while Nvidia lost more than 9% and nearly $450 billion in market value.

Elon Musk says SpaceX is adding two sites, called Gun Range and Spam, to its Memphis-area Colossus 2 computing complex; Data Center Dynamics reports that both are in Southaven, Mississippi.

OpenAI published a voluntary community compact for future compute campuses it owns and operates, plus a policy blueprint urging governments to set broader rules.

OpenAI CEO Sam Altman responded to criticism of his comments about DRAM in personal computers by saying he wants people to have as much memory as they want and that production should expand.

Super Micro Computer reported fiscal first-quarter revenue of $19.95 billion and raised its fiscal 2027 revenue outlook to $72 billion–$82 billion.

CoreWeave says most of a planned 22.5 MW deployment is now online at its first Danish data center, while it begins extending an AI-focused network backbone into the Middle East and Asia Pacific.

humans& describes a lossless GPU-accelerated pipeline for synchronizing reinforcement-learning weights between trainers and rollout samplers.

Five major data-center builders have disclosed more than $3 trillion in commitments tied to buildings, leases, energy and equipment over the coming years, Bloomberg reports.

AMD says initial Helios rack-scale systems are installed and being tested at Meta, expanding a partnership that spans CPUs, GPUs, networking and software.

Alphabet, Amazon, Meta and Microsoft have outlined more than $740 billion in capital expenditure for the year, spending investors largely read as AI infrastructure investment.

Google’s Gemini Notebook Android app can record a voice note up to one hour long and add its transcript as a notebook source; it does not retain the original recording.

Two Carnegie Mellon University researchers argue that computer-use agents can serve as embodied generalists when physical tasks are exposed through software interfaces.

Google’s Computer Use tool for the Gemini API is now generally available, with no breaking changes from preview.

Playcode says Gemini 4 Argon’s introductory price on its platform ended earlier than promised, doubling the listed rate to $4 per million input tokens and $20 per million output tokens.

Google’s Gemini 3.8 Flash TTS documentation lists Voice Remixing as available through both the Gemini API and Google AI Studio.

Vals AI compared three reasoning settings across seven frontier models and eight benchmarks, finding that lower settings often cut run cost much more than they reduce benchmark score.

Xiaomi says its humanoid robots completed an eight-hour sorting shift at a workstation in its electric-vehicle factory, detecting and correcting anomalies and coordinating with the production line.

Collinear AI has released CWE-bench v2, a public benchmark of 100 tasks across 91 vulnerability classes for agents that find and fix security weaknesses in real open-source repositories.

Google says students can now add hundreds of textbooks from providers in the U.S.

Google Cloud’s documentation describes Gemini Robotics ER 2, a preview vision-language model for reasoning about physical environments and coordinating robot tasks.

Vectorize released three open-weight models for its Hindsight agent-memory system and a hosted Inference API compatible with Cohere and OpenAI clients.

A Voximplant engineering article describes using fast decision classifiers such as TypeSafe’s Jev alongside an LLM in production voice bots.

Hume has introduced an Audio Decision Model API that accepts audio plus developer-defined questions and returns probabilities for possible answers.

Beget says a cyberattack overnight on 2 November destroyed its cloud infrastructure, including VPS, and that file backups will be available for only about half of VPS users.

Abdulrahman Alfozan, a former OpenAI engineer, pleaded not guilty on November 6 to one federal count of intercepting oral communications.

The Go Security team says it shipped Go 1.27.3 and 1.26.10 within about seven and a half hours of learning that weak eMudhra certificate keys could enable attackers to impersonate TLS servers.

Cloudflare’s November 6, 2026, technical post examines how Privacy Pass-style abuse controls could withstand quantum attackers while preserving unlinkability.

Cloudflare says Workers Builds and Pages build images now reject certificates chaining to eight compromised eMudhra intermediate authorities.

Binance began restoring withdrawals in phases after more than $6 billion in virtual assets were stolen, while reporting $3.64 billion in net outflows during a Sunday window.

OpenAI says an autonomous pipeline found and dynamically reproduced Linux kernel vulnerabilities, generated 948 working local privilege-escalation exploits, and prepared candidate patches in a mid-October run costing about $1.21 million in API credits.

Beget says attackers accessed management servers, ran malware intended to delete data, and prompted the company to disconnect all servers on November 2, 2026.

NCSC-NL’s php2shell repository now tracks 948 products potentially affected by two PHP remote-code-execution vulnerabilities, with 215 listed as vulnerable and 185 as fixed in its 4 November 2026 snapshot.

Cloudflare has introduced Zero Wire, a hub on Cloudflare Radar that combines plain-language briefings on critical vulnerabilities, global attack telemetry, and remediation steps.

Summary OpenAI says former U.S. Army Secretary Dan Driscoll will join the company in a newly created national-security advisory role. The appointment signals an effort to connect advanced AI with government and allied institutions but the announcement describes responsibilities and

A cyberattack overnight on November 2 disrupted refrigeration in supermarket stores in Sweden, Belgium and Luxembourg, with one Swedish store reporting freezer temperatures of 50°C and ruined stock.

Michigan’s Department of Corrections will pilot a real-time software module that assigns incarcerated people security classifications based on behavior and needs.

Summary Amazon is beginning an invite-only U.S. rollout of text messaging with Alexa for Shopping an agent that can search compare assemble carts and make purchases. The move brings an existing shopping agent into SMS while preserving continuity with Amazon’s

Datasette Apps 0.3a0 adds permission-aware app references and stored-query controls to workflows using Datasette Agent.

AWS describes Amazon Quick through four questions—when work starts, what it must do, who does it, and how it is done—and maps them to triggers, agents, and capabilities.

OpenAI’s Agents API reference documents an experimental time_awareness setting that gives an agent periodic reminders of the current time.

ElevenLabs says Eleven Realtime is now available to ElevenAgents voice agents in every workspace, completing a rollout that began on October 19.

Standard Physics presents Reynolds as an agent that turns a plain-language flow description into an OpenFOAM simulation and returns figures, numerical results and a rerunnable case.

Engram reports that reinforcement learning on top of fixed text notes improved agent accuracy by 7–20 percentage points across three workspaces.

OpenClaw v2026.11.1 is a patch release, not a feature release.

DR Pixel is an open-source deep-research agent that retrieves rendered document pages and reasons over page images rather than converting papers into OCR text first.

Broccoli AI has raised $52 million across two venture rounds to build agents for plumbing, HVAC, electrical, and other home-service businesses.

A team reports an autonomous research workflow in which AI agents formulated and tested predictions about quantum spins in diamond.

Tencent Hunyuan has released Hyra 2.0, an AutoResearch agent presented as a system for producing solutions and reusable artifacts across open problems in science, mathematics, engineering, and creative design.

Oakland police are investigating the disappearance of six pallets of Nvidia hardware valued at $11.4 million while en route to Fort Worth, Texas.

A November 3, 2026 arXiv preprint presents ForgeKernel, a GPU-kernel generation system that carries validated architectural knowledge across optimization runs.

Uber benchmarked Onehouse Quanton, a commercial Spark engine, against open-source Spark and Hudi on two synthetic, production-shaped table workloads.

SemiAnalysis says SpaceX has raised its 2027 AI datacenter capacity target to about 13 GW, with roughly 3.6 GW under construction and more than 9 GW of sites identified.

San Francisco’s mayor approved an urgency ordinance imposing a 45-day citywide zoning moratorium on data centers as defined by the measure.

Taiwanese lawmakers have proposed criminal penalties for unapproved exports of specified advanced chips and related equipment to China, including shipments routed through third countries.

Linum says it trained an experimental 2.18-billion-parameter Pyramid-JiT image model on 660 million samples for $21,000 in compute.

Apple is beta-testing four Audio Intelligence features for Apple Watch Series 12 and Ultra 4, including short conversation replay and summaries.

Core Automation reports training a roughly 32-billion-parameter sparse mixture-of-experts model across two GPU clusters about 9,000 miles apart, using different accelerators and parallelism schemes.

OpenAI has launched Grid Protectors, a program to coordinate future compute campuses with utilities so they can reduce electricity use when local grids are strained.

Visa launched an open benchmark that tests whether shopping agents can uncover a buyer’s requirements and make the right purchase-or-no-purchase decision.

A research project called STAND argues that a handful of archived industrial failure episodes can make a lightweight supervised anomaly detector more effective than larger unsupervised systems on the benchmarks tested.

An independent ShadowBench baseline shows that rigidly moving open hands can earn substantial silhouette-IoU scores without producing convincing target figures.

OpenAI and biotech company Ligo have published BioBasicsBench, a 100-task benchmark for testing AI agents on evidence from everyday wet-lab work.

A research team introduces ReproBridge, an agent workflow that turns a paper and its official code into a refined, reusable reproducibility artifact.

RoboColosseum says it is building a real-world evaluation ground for robot policies, spanning more than 100 robots and 140 scenes.

MindStudio’s November 1, 2026 article explains how training against public benchmark questions or formats can inflate a model’s leaderboard result without establishing broad capability.

A team introduced BRACE, a framework for generating training examples that test whether AI claims about chest radiographs are supported both by a report and by cited evidence.

LangChain argues that agent evaluation improves when domain experts review full conversations, explain their judgments and turn those judgments into reusable evaluation data.

Andreessen Horowitz says it is leading a $75 million Series B in Tutor Intelligence, a robotics company deploying remotely supervised robots in manufacturing and logistics.

China’s state planner has prohibited SoftBank’s $5.4 billion acquisition of ABB’s robotics business, citing concerns about overseas transfer of robotics data, Reuters reports.

Waymo says riders in Austin and Atlanta will be able to book its robotaxis directly through the company’s own app starting in January 2028, ending its Uber exclusivity in those markets.

Vercel has launched a public beta of Schedules, a service for static and dynamic scheduled work.

Serve Robotics and DoorDash have launched an opt-in footpath-robot delivery pilot in the Gold Coast Health and Knowledge Precinct, with Gnocchi Gnocchi Brothers as the first participating restaurant.

Usagi-org’s open-source ai-goofish-monitor is a self-hosted tool for monitoring listings on Xianyu, the Chinese second-hand marketplace.

A research team describes a quadruped locomotion pipeline that combines proprioceptive terrain estimation with safety-critical control.

Foundry Robotics raised an oversubscribed $100 million Series A to expand its California assembly facility, develop robotics and hire.

Walmart and Symbotic have expanded their commercial agreement to add automation capabilities across 25 U.S.

Meta is rolling out real-time voice conversations and expressive avatars in its Muse app in the United States.

VFlash is a research method for generating visual plans as whole structured primitives rather than predicting coordinates one token at a time.

Subquadratic released SubQ Snap 1.1, a post-training update to its decision model that retains a 384K-token context window while reducing input pricing by 20%.

Zed’s October 2026 preview release 1.27 reduces work performed by its Agent Panel when opening long threads and while an agent is active.

OpenAI says ChatGPT Voice has arrived in Finances, letting users connect financial accounts and discuss spending, savings and other money questions aloud.

OpenClaw v2026.10.10 adds a Details panel that brings session information and task progress beside wide conversations, plus an experimental, opt-in Advisor that can suggest course corrections during a task.

Sarvam has released Prism, an Apache 2.0 decision model built on Sarvam-105B that scores fixed answer options instead of generating free-form text.

Octen has published Octen-WebGR-4B, a generative retrieval model for navigational search that maps queries to semantic identifiers for URLs.

Amazon engineering leadership has updated its software development engineer role guideline to account for AI while keeping customer, business, and team outcomes at its center.

Moonshot has launched Kimi Hosted Agents, a managed cloud runtime for long-running and multi-agent workflows.

Perplexity is introducing Teammates, persistent work agents available on the web to Perplexity Max subscribers.

Shopify describes Reflex, an architecture for building and testing native mobile apps with agents.

Alibaba researchers describe SPAR, a framework for multi-session B2B sales agents tested in a live A/B test on the 1688 platform with more than 200,000 merchants.

Standard Physics says its Archimedes system uses agents to write engineering designs as code, test them in physics simulations and iteratively improve the strongest candidates.

Deloitte’s October 2026 State of AI in the Enterprise report finds that AI is widely associated with faster decisions, while deep business transformation remains much less common.

An ACM Queue article proposes a holistic way to discuss the cost and value of commercial software development across development, deployment, and operation.

Salesforce has signed a definitive agreement to acquire Cresta, whose voice agents handle complex customer calls and whose platform also coaches human representatives.

Taxbit says its embedded Assist agent has been in production in the US and EU since May 2026, answering dashboard questions with cited product knowledge and role-aware workflow checks.

Mistral’s documentation places Managed Agents and the Agents SDK in public preview, while Workflows are generally available.

Mistral has expanded Studio with managed agents and apps, alongside workflow, enterprise-context, observability and evaluation capabilities.

Switch presents itself as an open-source platform for placing AI teammates in Slack, Microsoft Teams, Discord, Telegram, and Mattermost.

A paper on Era by Eon reports that 918 of 2,304 answer-producing runs in a simulated enterprise benchmark contained a policy violation.

Circle K says an agentic system built with Amazon Bedrock reduced its mean DevOps incident-triage time to under 15 minutes, from a prior 45–60 minutes.

OpenAI is replacing maximum-bid controls with average-cost targets for new ChatGPT Ads campaigns beginning November 9, 2026.

Agility Robotics says it has completed its business combination with Churchill Capital Corp XI, with the combined company expected to begin Nasdaq trading as AGLT on November 5, 2026.

Character Technologies’ latest proposed sale procedures keep Disney as the lead bidder, retain a $700,000 break-up fee and a $21.7 million floor for competing bids, but shift key deadlines into December.

Thirty-one states and the District of Columbia have objected to the proposed bidding rules in Character.AI’s Delaware bankruptcy sale, arguing that the structure could deter offers that would leave the estate with more money than Disney’s $20 million deal.

Idealab founder Bill Gross has opened a free catalog of more than 1.2 million AI-generated business ideas, built by applying his 31 idea-generation methodologies to roughly 17 million source documents.

Microsoft reported $92.3 billion in revenue for the quarter ended September 30, 2026, with Microsoft Cloud revenue up 30% and Azure up 48%.

In a Stratechery interview, Palantir CTO Shyam Sankar says that almost all customers he has worked with who wanted AI to reduce headcount failed to meet that goal.

Proton CTO Bart Butler published an internal engineering memo arguing that AI changes engineers’ work from manually writing code to directing models and reshaping delivery workflows.

A 404 Media investigation describes Proton leadership pressing engineers to use AI more aggressively while employees raise concerns about security, workplace pressure, and the pace of deployment.

Summary The U.S. Census Bureau changed the language of its business AI-use survey to Super Intelligence or SI following Executive Order 14434. The shift follows an earlier broadening of the question breaking direct comparison across survey periods. For anyone interpreting

State of AI Report compares three widely cited 2026 measures of business AI adoption and shows why they do not describe a single rate.

A Habr case study describes a five-layer memory system behind an assistant used by about 1,200 people across Telegram, MAX, and VK.

MIT researchers describe UQ-DAAAM, a system that identifies uncertain object descriptions in a robot’s map and revises them using views captured from other angles.

Stanford’s MEMORILLA learns to read an entire document collection and turn each query into a fixed set of memory tokens for a frozen language model.

Vercel Labs’ fx v0.0.15 reports lower memory use in long conversations and adds session branching with /fork.

Alex Wang’s essay recasts linear maps, MLPs and self-attention as associative memory stores and estimates how many key-value associations each can hold under a stated interference measure.

The Browser Company will shut down Arc on December 22, 2026, and is directing users to its AI-first browser, Dia.

A research team presents HiRO-ACEv2, a two-stage AI emulator of the 3 km X-SHiELD global storm-resolving model.

Meta says Ray-Ban Meta Audio, its first audio glasses in the Ray-Ban AI-glasses collection, are now on sale from $349.

A USC-led analysis of 58,746 NIH-funded biomedical projects from 2025 estimates that AI accounts for 15.9% of the portfolio, with a 13.4% funding premium.

Scrapling v0.5.1 adds motion-aware browser observations, drag-and-drop control, and dedicated sessions for connecting to existing Chromium browsers.

Floci is a free, open-source emulator for AWS services that runs locally and accepts familiar AWS clients and infrastructure tools.

On November 4, 2026, LlamaIndex announced that LlamaParse no longer applies usage caps to its Pay-as-you-go and Pro plans, and published self-serve credit rates for both.

Scribe is a desktop LaTeX editor that brings a research paper, its Python notebooks, and generated figures into one project, compiled locally.

Nous Research’s official hermes-security plugin brings repository security review into Hermes Agent, with parallel reviewers, evidence-linked findings, optional test confirmation, and exportable reports.

A community plugin for Hermes Agent snapshots files before selected tool calls and offers commands to inspect and restore those changes.

Nous Research says it has released Arete-27B, an open-weight fine-tune of Alibaba’s Qwen3.8-27B intended to change how the model handles politically sensitive topics.

IBM Research compared CUGA, Hermes, and LangChain’s Deep Agents on the same open-weight model and 90 AppWorld tasks.

ModelBest says Lookahead Expert Scheduling (LES) lets MiniCPM-V 4.7, a 35-billion-parameter MoE model, decode above 20 tokens per second on an iPhone 18 Pro Max without changing the model's routing or network structure.

Ollama’s v0.40.11 release adds speech generation on macOS.

Prism ML has released Ternary-Bonsai-2-2B-gguf, a 2B-class vision-language model with ternary language weights in several compact GGUF packings.

Underdog's Hugging Face model card describes Woof as a small English text-to-speech model, now available there in a two-voice build that runs on a laptop CPU using NumPy.

Mistral’s changelog says Mistral Large 4 is in public preview as an open-weight, general-purpose multimodal Mixture-of-Experts model with a one-million-token context window.

Kakao has released Kanana-2-v-decide, an open-source multimodal model that returns probabilities, choices or scores instead of generated prose.

Meituan has introduced LongCat-2.5, a 1.6-trillion-parameter model with about 48 billion parameters active per token, aimed at long-horizon agentic work and native multimodal tasks.

Randall Balestriero and Yann LeCun present a theorem and experiments arguing that cross-modal representations can align more efficiently when training pairs show the same scene from different views.

NVIDIA says its open-source Aerial and Sionna tools are being used in 5G/6G research, with reported results spanning faster robot tasks, improved wireless reception, lower radio-access-network power use and quicker simulations.

The Linux Foundation says generative AI is helping open-source contributors complete development tasks faster, while maintainers face more incoming contributions and substantial low-quality pull-request volume.

OpenRoboKit’s RoboKit repository describes a general-purpose toolkit for parallel robotic optimization and learning, with support for differentiable kinematics, task objectives and optimization methods.

AGIBOT has announced GO-3 Preview, an embodied foundation model intended to learn across pre-training, post-training and deployment.

NousResearch has published Propaganda_Eval on Hugging Face: 803 structured records for examining answers to politically charged prompts.

JailBound is a jailbreak-evaluation framework that generates prompts from a structured threat taxonomy and searches for behavior-sensitive model states.

A Findings of EMNLP 2026 paper studies how prompt design, examples, model choice and workflow affect LLM-based machine-translation evaluation.

Project Maya v1.1.35 changes how GLM experts are kept on NVIDIA GPUs: when the full expert set fits in VRAM, it stays resident through long prompts rather than being evicted.

Pinecone Terraform Provider v5.1.0 makes on-demand read capacity the default for newly created Bring Your Own Cloud indexes and adds integrated embedding to document-index string fields.

A public repository demonstrates an end-to-end migration of a Milvus 3.0.2 collection from a smaller to a larger BGE embedding model, retaining the old vector field through backfill, evaluation, canary routing and rollback.

Milvus says Vector Graph RAG now uses Jev in two retrieval stages: first to select graph relationships, then to rerank source passages alongside direct vector-search results.

Milvus compares four HNSW index variants in version 2.6.17, focusing on how vector compression and optional reranking affect retrieval trade-offs.

Rosselkhozbank describes replacing a vendor chatbot platform with an internal system, then improving intent recognition using a compact multilingual MiniLM model that runs on CPU.

Supermicro, IBM and NVIDIA announced a jointly validated system that stores reusable inference KV states in shared network storage rather than relying only on GPU or CPU memory.

SpaceXSI says its data science team built Data Bot, a shared Grok Bot that queries the company warehouse, runs analyses and can support recurring workflows in Slack.

Paul Ford argues that modern browser capabilities, paired with lower development and hosting costs, make ambitious software projects more accessible to individuals.

Discord is rebuilding its desktop client in Rust and using GPUI for its interface, replacing Electron; the mobile app will keep a native UI while sharing code with desktop through Rust.

NanoUI is an open-source Zig toolkit for building small desktop apps with either GPU-drawn native UI or a web frontend.

Meta says Muse users can now create custom interactive “Spaces” by describing an idea, then revise and organize those experiences inside the app.

Airbnb says it plans to begin rolling out a specialized AI travel agent to guests in the first half of 2027.

A paper by Hita Kambhamettu, Alyssa Hwang, Philippe Laban and Andrew Head proposes “attribution gradients,” an interface for inspecting the evidence behind citations in AI-generated answers.

Tesla says Optimus production has begun on high-volume lines at its Fremont factory, and calls the robots from that line its first production-quality bots.

Together AI says its Together Compiler (TKCC) uses parallel agents and persistent inference engines to keep testing workload-specific optimizations without repeatedly restarting engines.

Factory says HumanLayer is joining the company, and that the existing HumanLayer platform will remain online during a transition.

Vercel has made Token Pass credits apply automatically to eligible model requests sent through AI Gateway using a standard API key, OIDC token, or fx login.

Baseten announced Grid for model routing and governance, Sandboxes for agent execution, and general availability for its serverless training product, Loops.

At its November 3, 2026 Forge conference, Fireworks announced products spanning inference, model routing, training and enterprise deployment.

Langfuse says Skill Management is now available in beta, with collaboration and versioning for AI agent skills and a way to connect them to agent traces.

DoorDash compared self-distillation and GRPO for sequentially teaching Nemotron Nano 9B new skills while tracking performance on existing capabilities.

Google says its new Gemini 4 Carbon model improves coding, agentic work and cybersecurity, with access initially limited to cyber defenders and trusted partners.

Mistral says its improved Large 4 model is now available with open weights, targeting agentic coding and defensive cybersecurity.

OpenAI is inviting a small number of software companies to bring difficult computer-use workflows into a research collaboration.

Tencent announced Hy4, a new-generation language model that adds visual understanding and computer use to the Hy line.

OpenViking documents a memory plugin for DeepSeek Harness that carries relevant information across projects and sessions, while capturing new content during conversations.

LangChain CEO Harrison Chase argues that proactive, always-on agents introduce context-engineering problems beyond the now-standard toolkit for interactive agents.

IBM reports that AI-related incidents reached the board or C-suite at 80% of surveyed EMEA organizations in the past year, while only 40% changed governance policy in response.

Open Data Agent Platform has released ODAP Gateway v0.2.0-preview.1, a pre-release describing tenant-scoped semantic review, scoped access requests, audit evidence and fail-closed authentication.

Russia’s digital development ministry helped Yandex find backup sites and move key systems after drone strikes hit three of its data centers.

Yandex says users can again access 96% of the data stored on Yandex Disk, though recovery varies by account and files are still being returned gradually.

AWS’s second Okta Cross App Access and Amazon Bedrock AgentCore guide moves from a small agent-and-tool demonstration to a multi-agent access pattern.

Researchers describe FuzzMedic, an agent system for repairing build failures that leave OSS-Fuzz targets inactive.

A NeurIPS 2026 paper reports that fabricated safety warnings embedded in ordinary reviews can steer agents without issuing explicit commands.

A new paper introduces Heads of Security (HoS), a model-architecture defense that learns attention gates to separate trusted instructions from untrusted content.

Telegram has published widget-sandbox, an open-source framework for running small HTML widgets inside Telegram messages.

Vercel says Sandboxes that mount Drives now start faster: median creation time fell from 1.7 seconds to 0.7 seconds, while p95 fell from 2.2 seconds to 1.6 seconds.

A preregistered study of 73 API and runtime migration tasks found that ordinary tests on typical inputs released 41% of incorrect LLM repairs.

Axiom Math’s 28 October 2026 document records 96 selected departures found while formalizing the classification of finite simple groups in Lean.

AWS describes a reference implementation for governing agents that act on a user’s behalf across applications, combining Okta Cross App Access with Amazon Bedrock AgentCore Identity and runtime.

Yohei Nakajima’s babyagi-inside repository describes a bounded BabyAGI-style loop compiled into widened Qwen3.5-2B checkpoints.

Relinos is an open-source programmable browser exposed through a REST API for rendering pages, interacting with sites and extracting structured data.

CrowdStrike announced agreements to acquire Fiddler AI and Dreadnode, adding agent observability and automated AI security testing to its platform.

Meta has open-sourced Astryx, a customizable React design system that it says grew inside the company over eight years and now powers more than 13,000 apps.

OpenAI introduced GDPval, an evaluation of model performance on realistic knowledge-work tasks across 44 occupations and nine U.S.

Vercel says Tracing is now generally available on every plan, with dashboard and CLI search, longer retention, and automatic tracing for qualifying new Next.js projects.

Auctopus 1.0.0 is an MIT-licensed Python package and command-line tool for orchestrating AI agents in experimental research.

OpenAI has added a cap on how many paid ChatGPT accounts can be used on one device, according to The Decoder’s report on the company’s support page.

Ted Tschopp’s essay turns human review into a capacity-planning problem for teams scaling AI workflows.

NASA’s Jet Propulsion Laboratory describes Teghaza 001, an 800-gram meteorite found in Mali, as one of the oldest Martian rocks recovered on Earth.

Bambu Lab announced the H3C, a desktop FFF 3D printer built around nine independently heated print tools and a material switch it says takes under six seconds.

X has added a credited image-sharing flow that lets a user repost another account’s photo while retaining a visible link to the original creator.

Rest of World, a nonprofit technology publication, plans to stop publishing independently at the end of January unless it finds a new institutional home.

The FCC’s adopted order ends a decades-old airborne-use ban for drones in two 800 MHz bands, while a new footnote names SpaceX’s pending bid for adjacent spectrum.

Science Corporation says it implanted its PRIMA retinal prosthesis in its first patient with Stargardt disease during a trial in Australia.